How to Finance a GPU Cluster

GPU cluster financing resembles a repeatable formula: raise 30-40% in equity (investor capital) borrow 70% debt (lender capital) through a special purpose vehicle (SPV). At the $5M-$100M range, GPU-backed term loans from small boutique lenders and private credit funds are the standard instrument. The three variables that determine your terms are: deal structure, projected utilization, and how you layer equity, senior debt, and mezzanine to close the gap.

The formula

Every GPU cluster deal at the sub-$100M scale, whether you're a neocloud, an independent compute provider, or a startup building inference infrastructure, follows the same pattern:

1. The operator forms an SPV.
2. The operator raises equity.
3. The operator procures hardware and data center space with a downpayment.
4. The operator secures compute offtake contracts (agreements where customers commit to purchasing a set amount of compute over a fixed period).
5. The operator borrows for the rest of the hardware cost using the SPV's assets and offtake contracts as collateral.

This is the same form of financing that has become popular in the past 20 years for other private credit deals. CoreWeave raised $7.5 billion in debt from Blackstone in May 2024 using this structure. ^{[1]CNBC, "AI startup CoreWeave raises $7.5 billion in debt, Blackstone leads" (May 2024)https://www.cnbc.com/2024/05/17/ai-startup-coreweave-raises-7point5-billion-in-debt-blackstone-leads.html}

The chicken-and-egg problem

How much equity you need depends on how much debt you can raise. How much debt you can raise depends on your offtake contracts. And customers only sign offtake contracts if they take your project seriously.

You solve this by getting soft commitments from all sides before hardening them. LOIs (letters of intent) or term sheets from customers, lenders, and equity investors.

Raising the equity

Equity is the first-loss dollar in the financing stack. Whether it's cash supplied by the operator, or investments from investors, it has no collateral protection. If the GPU cluster goes bankrupt, equity doesn't get paid back from the sale of any collateral. It only earns any profits from the project after the debt is repaid; giving it a longer payback horizon than lenders.

But equity gets the highest potential returns, even if it's not fixed. Sources for equity at the $5M-$100M scale include: private equity, family offices, high-net-worth individuals, or strategic investors (who plan to use the GPUs for their own business). Lambda raised $480 million in equity in February 2025, then another $1.5 billion in November 2025, in part to support the debt capacity needed for its GPU fleet expansion. ^{[2]CNBC, "AI cloud startup Lambda raises $480 million in new round, Nvidia among investors" (February 2025)https://www.cnbc.com/2025/02/19/ai-cloud-startup-lambda-raises-480-million-in-new-round-nvidia-among-investors.html [3]Reuters, "Lambda raises $1.5 billion in equity financing at $4.4 billion valuation" (November 2025)https://www.reuters.com/technology/lambda-raises-15-billion-equity-financing-44-billion-valuation-2025-11-20/}

At 80% utilization (% of time the cluster is generating revenue) on a 1,024-GPU cluster, equity investors can likely see 2-3x MOIC (multiple on invested capital) over 5 years based on current market rates. At 60% utilization, they'll more or less break even. Below 60%, interest payments and operating costs exceed the revenue the cluster generates, and equity investors absorb those losses first. This means below 60%, equity investors will lose money.

Strategic equity

A customer can put in equity in exchange for reserved capacity at a discount. This aligns incentives. If things are delayed, the customer is more patient because they're an investor and want the project to succeed. It also gives the lender confidence with a built-in offtake commitment.

How much equity is enough

70-30 LTV (loan-to-value, the ratio of the loan amount to the hardware cost) is common. The lender wants to see the operator have 30% equity and 70% debt. Operators with existing clusters and operating history can push to 75-25 LTV or more. First-time operators with no track record may need 40%+ equity.

The equity serves as a buffer for the lender. If the operator can't pay back the loan, equity investors absorb losses first. The lender only loses money once all the equity is wiped out. More equity means more cushion before the lender is exposed.

The equity raised isn't just for the hardware downpayment. It also needs to cover operating losses until the cluster is cash-flow positive. Loan covenants strictly prohibit the loan from being used for non-hardware expenses such as salaries, admin, insurance, and software. Lenders enforce this because operating expenses are hard to audit and don't produce seizable collateral; they want every dollar of the loan going into revenue-generating hardware they can repossess.

For a 1,024-GPU B200 cluster at 70% LTV, the hardware equity is roughly $15M (30% of the $47M capex, plus a 3% origination fee, the one-time cost the lender charges to underwrite and close the deal). But our modeling shows that total cash needed to launch, is closer to $16.5M. The extra $1.5M is due to the fact that the cluster needs to cover operating losses until it is cash-flow positive.

How to structure the deal

A special purpose vehicle (SPV) or special purpose entity (SPE) is a standalone legal entity that holds the GPU hardware, the compute contracts, and the debt.

It isolates the cluster from the operator's other liabilities and businesses. If the operator goes bankrupt, the lender's claim on the SPV assets is protected. CoreWeave uses this structure through entities like Compute Acquisition Co. IV, LLC, which is a Delaware limited liability company. ^{[4]SEC EDGAR, CoreWeave Compute Acquisition Co. IV, LLC credit agreement filings (2024-2025)https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798dex1013.htm}

The lender has a security interest in the GPUs (a UCC Article 9 filing in the U.S.). ^{[5]Cornell Law School, Legal Information Institute, "U.C.C. Article 9 - Secured Transactions"https://www.law.cornell.edu/ucc/9} Each GPU is identifiable by serial number, and the lender records a lien (ability to seize the hardware) against all hardware in the SPV. The offtake contracts are assigned to the lender as collateral as well. If payments stop, the lender can seize the hardware and resell the customer contracts. ^{[6]Bird & Bird / Lexology, "GPU-Based Financing in the Global Data Center Market" (2025)https://www.lexology.com/library/detail.aspx?g=71bf28ab-ce78-46ba-be78-9c9a09464767}

Delayed draw term loans

Hardware arrives in tranches, not all at once. A DDTL (delayed draw term loan) lets you draw capital as servers arrive and get racked. You do not pay interest on capital you have not drawn yet. Nscale signed a $1.4 billion DDTL in February 2026 structured this way. ^{[7]Nscale, "$1.4bn Delayed Draw Term Loan Backed by GPUs" press release (February 2026)https://www.nscale.com/press-releases/nscale-signs-1-4bn-delayed-draw-term-loan [8]Proskauer Rose LLP, "Private Credit Explained: Delayed Draw Term Loans" (2024)https://www.proskauer.com/alert/private-credit-explained-delayed-draw-term-loans}

For a 1,024-GPU cluster, hardware might arrive in 3 tranches over several months. You draw capital as each batch arrives and gets racked. This reduces the interest cost dramatically compared to drawing everything upfront.

Bridge financing

Lenders sometimes require bridge financing to cover the 90 days between hardware payment and installation. A bridge lender pays for the hardware. Once the servers are racked and operational, the main lender pays off the bridge lender.

Bridge loans carry higher interest rates than the main term loan because they are higher risk. If the installation fails, the main lender can refuse to pay off the bridge lender. For operators who cannot tie up equity in hardware deposits for months while waiting on deployment, bridge financing can keep a project moving.

What lenders require

Lenders want offtake contracts covering at least 1x debt service (the monthly repayments of the loan's principal and interest). They want a deployment timeline with milestones, and audited financials if the borrower is established, or detailed projections if not. They'll also want insurance on the hardware.

Most importantly, the lender will want to see that monthly GPU revenue exceeds monthly debt service by at least 1.2-1.5x, the debt service coverage ratio (DSCR). If your revenue is not enough to cover the debt service, and deployment timelines seem hard to hit (especially if you have a limited track record) the lender will not move ahead with the loan. Tools like residual value insurance can help get them over the line, but it's entirely deal dependent.

Interest rates for GPU-backed term loans are typically 15% for less experienced operators, but lower for operators with a track record and contracted revenue. Loans are written with simple amortization (equal payments over the term), with periods of 3 to 5 years.

OEM financing

Many OEMs have financing arms that can provide equipment loans directly. Dell has Dell Financial Services, Cisco has Cisco Capital, HPE has HPE Financial Services, and Lenovo has Lenovo Financial Services. These programs can offer competitive rates because the OEM already understands the hardware value. They can repossess, refurbish, and resell it through their own certified pre-owned channels if the borrower defaults.

The catch is that smaller operators often struggle to get on their radar and pass their diligence requirements, but for those who qualify it can be a straightforward path to financing.

The deal timeline

A GPU cluster deal is not sequential. Most workstreams run in parallel.

Six workstreams run in parallel from month zero. Equity and offtake start simultaneously because each needs the other: investors want to see customer demand, customers want to see that the operator can deliver.

Colocation and procurement start within weeks, using the initial equity raised as an equity downpayment to place hardware orders and sign a lease. Both need to be settled early enough or the deployment window slips.

Debt closes after equity is committed and offtake contracts are signed.

How mezzanine fills the gap

Imagine you're raising $100M. Lenders cap at 60-70% LTV, meaning they'll only lend 60-70% of the total hardware cost. Equity investors want to minimize their check, and you can only get 10% raised in equity. How do you fill the gap?

Mezzanine fills the remaining 10-20%. Mezzanine debt, sometimes called subordinated debt or junior debt, sits between the main lender (senior debt) and equity in the capital stack. In a default scenario where the operator goes bankrupt, the senior lender gets paid first, then the mezzanine lender, second, and equity last. The higher risk means higher return: 15-20% interest for mezzanine versus 12-15% for senior debt. ^{[9]Wall Street Prep, "Mezzanine Financing: Definition and Debt Characteristics"https://www.wallstreetprep.com/knowledge/mezzanine-financing/}

When mezzanine does not make sense

If total debt service (monthly loan payments of the senior + mezzanine) exceeds the revenue the cluster generates, the deal is not sustainable. The mezzanine tranche (portion of the financing stack) only works when the operator has enough margin at realistic utilization rates to service both tranches of debt and still cover operating costs.

Mezzanine lenders often behave as venture debt, and require warrants. Warrants give them the right to convert some of their debt into equity if the company is doing well. They control the option and can exercise it only when it benefits them. A 10% warrant on a $7M mezzanine tranche can end up costing more than the interest payments if the cluster is doing well.

Why utilization determines everything

Lenders underwrite the revenue stream, as much as if not more than the hardware. GPU clusters have high fixed costs: hardware payments, colocation fees, power minimums, insurance. The additional cost to run one more GPU-hour is close to zero. This makes utilization the single biggest lever on unit economics.

Breakeven for a financed cluster sits at roughly 60% utilization. Below that, the operator cannot cover debt service plus operating costs. Above that, the margin expands quickly. The difference between 55% and 85% utilization on a 1,024-GPU cluster is the difference between losing $330,000 per month and making $340,000 per month.

How lenders evaluate utilization

Lenders look at three things. First, contract backlog: committed revenue from signed compute agreements. Second, historical utilization on the borrower's existing clusters, if any. Third, demand letters or LOIs from prospective customers. A cluster with 60-70% of capacity under signed contracts is fundable.

Reserved instances, where the customer pays whether they use the GPUs or not, and take-or-pay contracts, where the customer commits to a minimum spend per month, are the preferred contract types for lenders.

On-demand revenue is real but lenders discount it heavily. A cluster with 100% on-demand has weaker contract coverage and will get lower LTV and higher interest rates.

What goes wrong

• Demand evaporates: If committed capacity gets cancelled or contracts are not renewed, the offtake backing the debt evaporates. This can happen when timelines slip, allowing customers to walk since most compute contracts have delivery deadlines. Lenders mitigate this by limiting customer concentration. A single customer representing more than 40% of revenue is a yellow flag. The lenders can also check the customer's creditworthiness and financials.
• The facility or hardware is not ready: If the site isn't ready on time, or the hardware doesn't arrive on time, the cluster cannot be deployed and generate revenue.
• Market prices compress: If market rates fall faster than expected, your existing reserved contracts are the lifeline. When those contracts expire, you may have to renegotiate at whatever the market will bear.
• Over-leverage: Stacking senior and mezzanine debt too aggressively means even small utilization drops blow through cash reserves.

References

1. CNBC, "AI startup CoreWeave raises $7.5 billion in debt, Blackstone leads" (May 2024)
2. CNBC, "AI cloud startup Lambda raises $480 million in new round, Nvidia among investors" (February 2025)
3. Reuters, "Lambda raises $1.5 billion in equity financing at $4.4 billion valuation" (November 2025)
4. SEC EDGAR, CoreWeave Compute Acquisition Co. IV, LLC credit agreement filings (2024-2025)
5. Cornell Law School, Legal Information Institute, "U.C.C. Article 9 - Secured Transactions"
6. Bird & Bird / Lexology, "GPU-Based Financing in the Global Data Center Market" (2025)
7. Nscale, "$1.4bn Delayed Draw Term Loan Backed by GPUs" press release (February 2026)
8. Proskauer Rose LLP, "Private Credit Explained: Delayed Draw Term Loans" (2024)
9. Wall Street Prep, "Mezzanine Financing: Definition and Debt Characteristics"